Schneier on Security

Ray Dillinger • December 8, 2025 2:36 PM

A few years ago someone claimed to have solved the Voynich MS and pestered the cryptography list with his claims. Basically his “decryption” of the MS is a mishmosh of words from a half-dozen languages thrown together in “sentences” which, although he ignores the fact, have no coherent grammar.

I wanted to at least know what I was talking about when I replied, so I went and gave it a semi-serious look.

Voynichese has a very strong bias about letters appearing cycles, in more-or-less the same order within each cycle – to the point where it seems like a lot of cycles come down to just one or two choices besides just deciding which majority of the “alphabet” is to be left out. It reminded me of an old fashioned stenogram, where the court reporter presses keys for a word on two chorded keyboards all at once, and the (approximate) vowels of a word, corresponding the left-hand keyboard appear on the left half of the page and the (approximate) consonants of the same word, corresponding to the right-hand keyboard, appear on the right, except that the Voynichese letter order wasn’t quite that completely regular. I could not detect any longer patterns which might have been about grammar rather than about individual words.

The looping sequence bias continues across spaces and line breaks, so whitespace appears not to be meaningful. When I talk about “words” I’m basically pointing at the loops in an almost-repeating cycle that runs, with variations, seven to fifteen letters per loop.

The sequence bias is represented fairly well (much better than a stenogram) with the tables of “prefixes” and “suffixes” that this paper uses to both encode information and bias the letter order within each loop. But the cipher uses spaces between iterations of the loop rather than spaces inserted more-or-less randomly as the MS does.

The most common letter in Voynichese is about two-thirds as frequent as English ‘E’ IIRC, but does not combine left or right with a wide variety of things the way vowels do. It does not appear to be used as a vowel in a substitution cipher. Neither, in fact, does any other letter in Voynichese appear to be used as a vowel in a substitution cipher. Each letter can appear after a very few things, and before a very few other things.

Finally the very strong sequence bias in Voynichese results in a ridiculous conditional probability table. If you see a sequence of some two letters, you can predict, with probability greater than 50%, the sequence of four or six letters to follow.

There are a set of letters bigger than others – you might guess that they are presentation forms like capital letters – which seem to appear only on the top line of each page and the rightmost character of other lines. But they don’t appear to be part of the repeating cycle substituting for anything else the way presentation forms would. This paper doesn’t attempt to reproduce their page positions, treating them simply as letters of somewhat more random placement and lower probability. So I strongly suspect this cipher must have been constructed from a “linearized” version of the manuscript, where the positions of page and line breaks wouldn’t be represented.

With all the highly regular statistical patterns based on sequence in the main loops, combined with a few letters that appear to follow a completely different but equally regular set of rules based on page position, it follows that if Voynichese represented a human language, it would have to be a very low-efficiency encoding – maybe around a third of a bit of entropy per letter. English writing by comparison achieves about one and a quarter or thereabouts. And that information density accords reasonably well with this cipher, which adds several random bits to each word by choosing randomly among encodings. Using mutally prefix-free Huffman encodings preserves the ability to decrypt while the randomness stretches out the meaningful entropy over more ciphertext characters. The mutually prefix-free Huffman encodings were a clever way to do this.

All told, it’s a nice piece of work, and keeps the hope of would-be decryptors alive.

That said, I’m still personally convinced that Voynichese is mainly or entirely meaningless. Probably the macguffin of a hoax somebody pulled way back in the wayback, or possibly the asemic product of someone’s untreated graphorrhea disorder in the same time frame. It’s not a modern hoax, but IMO it’s also probably not language.

The thing that argues against the latter (and therefore favors the ‘hoax’ theory) is the quality of the materials used. That is fine vellum and good mineral-pigment ink that would cost decades of a good craftsman’s earnings at the time, and a disordered mind in the throes of graphorrhea wouldn’t have access to such materials unless they were very wealthy or working for or with a very wealthy institution such as the Church. In which case the author, as a noble, royal, or other notable person having such a very specific disability, would likely be at least as well-known as Hildegard von Bingen. And no such person is known.